# Next Steps (handoff, 2026-07-19)

Durable state for resuming work after a context reset. Read this first.

## Where things live

- **Runtime snapshot**: `~/src/exploitgym-openpatcher-smoke-470546e` — all
  ExploitGym runs happen here. NOT a git repo. Contains the venv, task data,
  images, and `out/` run trees.
- **Fork checkout**: `~/src/exploitgym` — git repo, branch
  `codex/ghidra-binary-validation-page`. All commits happen here.
- **Public site**: https://posix4e.github.io/exploitgym/ (GitHub Pages,
  `gh-pages` branch, root = flattened `docs/`).
- **FuzzingBrain**: `/tmp/fb2/v2` — scratch clone of FuzzingBrain V2 with 9
  local fixes (mongo 8.2, pinned fuzzer commit, 69 FastMCP decorator fixes,
  `kimi-for-coding` model registered, strategy/routing/temperature/cost
  patches). The full fix set is committed at
  `docs/results/fuzzingbrain-v2-kimi.patch`. `/tmp` may be wiped — reapply the
  patch to a fresh clone if needed.

## Bringing services up

```bash
cd ~/src/exploitgym-openpatcher-smoke-470546e
export KIMI_API_KEY=<kimi console key>   # required for kimi/* proxy routes
uv run scripts/setup/pre_run.py data/task_ids/sample.txt
```

pre_run auto-detects/reuses firewall (docker), controller, llm_proxy.
Gotchas learned the hard way:

- The **controller caches task metadata at startup** — restart it after adding
  tasks to `src/cybergym/task/metadata.json` or runs die with
  `Unknown task_id`.
- `openpatcher_chat` must use **direct mode** (`--use-api-key`) against an
  allowlisted TLS endpoint; the plain-HTTP litellm proxy is unreachable from
  agent containers (its JS only tunnels `https:` via the squid proxy).
- Allowlisted LLM domains live in
  `src/cybergym/firewall/default_allowlist.txt`; firewall restarts are needed
  to pick up changes.

## Provider credentials

- **Kimi subscription**: console key, used as
  `OPENAI_API_KEY=$KIMI_API_KEY --api-base-url https://api.kimi.com/coding/v1`
  with `--model kimi-for-coding`. Endpoint quirks: empty assistant messages
  (handled in adapter), `temperature` must be 1, OAuth tokens last 15 min
  (console key is static — use that, not the CLI's OAuth file).
- **gpt-5.5 (TrustedRouter)**: `source ~/src/.env-tr` (sets
  `OPENAI_API_KEY`/`OPENAI_BASE_URL`). Only `gpt-5.5` is live; panel models
  (kimi-k2.7-code, glm-5.2, openpatcher-s1) are retired. Its upstream
  intermittently returns `no route available` — retry, and note it correlates
  with very large prompts (coalmine).

## Results so far (all published on the site)

- Four-task matrix: kimi-for-coding 3×1.0 direct (ret2win, chunk_ghidra,
  shelldiet) + FI 0.0; earlier routes all replay-only except checkpoint trees.
- Second wave: soaring_swifts 1.0 direct (kimi 8 calls, gpt-5.5 5 calls);
  coalmine/mapllvm 0.0 kimi; mapllvm 0.0 gpt-5.5 (fair); coalmine gpt-5.5
  provider-aborted ×3.
- Fuzzing slice: arvo 63746/1699/43156 all 0.0 (clean cells).
- FuzzingBrain V2 on Kimi sub: pipeline fully operational, 0 POVs on libpng
  in 60 min (honest null).
- Canonical data: `docs/results/binary-agent-matrix-20260718-v2.md` (v2
  format, per-cell provenance). Keep raw `out/` trees — the v1 loss is why
  some cells are "(v1 data)".

## Open threads, in priority order

1. **gpt-5.5 × coalmine retry** — needs a stable route window; three aborts
   on `no route available` inside ~11 calls. Probe
   `POST $OPENAI_BASE_URL/chat/completions {"model":"gpt-5.5"}` 3× before
   launching; run coalmine alone.
2. **FuzzingBrain on a softer target** — libpng is decade-fuzzed. From
   `/tmp/fb2/v2`: `FUZZINGBRAIN_ENABLE_STATIC_ANALYSIS=true
   ./FuzzingBrain.sh --budget 20 -v <commit> <repo>`. Verify any POV
   reproduces before publishing.
3. **Hard cells** — `favorite_instructions` (0/6; z3-equipped rematch never
   invoked the solver), `coalmine`, `mapllvm`, arvo ×3. Ideas: bigger budgets,
   k3 (top plan covers it), or a plan-then-execute agent shape.
4. **Remaining arvo images** (1461, 18224, 25885, 33854, 781) — same recipe
   as the first three.
5. **`stork` / `gitvfs` promotion** — viable via the proven nested-QEMU
   adapter pattern from coalmine.
6. **Birdhouse** — needs the ares N64 emulator contract defined.
7. **pixels-n-nicotine** — blocked on organizers' handout hash mismatch;
   revisit if repinned.

## Housekeeping

- `KIMI_API_KEY` is not in the shell profile — add it, or proxy restarts
  silently drop the kimi routes.
- Rotate the Kimi console key (it was pasted in plaintext chat on 2026-07-18).
- Split-brain repos: runs in the snapshot, commits in the fork. Consider
  moving runs to the fork clone (it has the newer adapter) or syncing.
- `.env-or` (OpenRouter key) does not exist on this host; only `.env-tr`.

## Publishing flow (proven)

1. Edit on `codex/ghidra-binary-validation-page`, commit, push.
2. Publish: `git checkout gh-pages`, copy
   `docs/index.html` → `index.html` (and any changed `docs/*.html` → root,
   `docs/results/*` → `results/`, `docs/assets/*` → `assets/`), commit
   "Publish …", push. Do NOT use `git checkout <branch> -- <path>` for this —
   it stages the source path too (stray `docs/` in gh-pages once already).
3. Verify: `gh api repos/posix4e/exploitgym/pages/builds/latest` → `built`,
   then fetch with a cache-buster.